This is the code belonging to the bachelors thesis RSLVQ as classifier layer in a convolutional neural network. The code was developed with Python 3; Python 2 might work, but it was not tested. Each method in the main files has a docstring to describe what it does; therefore, feel free to dive into the code :)

Part1.py contains the tests for part 1. The RSLVQ implementation used by it is found in rslvq1.py. It can run a single test, tests with the best 10 hyperparameters found in the thesis or a whole batch of hyperparameter tests. Shuffling the dataset is also possible. Run Part1.py --help to see all parameters. The --use-mp option uses the multiprocessing library instead of threads. When running many threads this parameter helps to improve the training time.

For it to work, the dog image dataset must be present (see CodeData/Instructions.md).

To run the comparison network execute Part1 comparison.py. This network needs the dogbreed dataset as well.

Part2.py contains the networks used in part 2 of the thesis.

Making it work is a bit more difficult: First, python3 setup.py build must be executed in the nnetcopy folder (this needs cython). This should create a build folder. The libraries from that folder (pool.so/pool.dll and conv.so/conv.dll) need to be copied to nnetcopy/nnet/convnet. After that Part2.py should be usable.

When starting the file without parameters, it will run the simple RSLVQ network on the MNIST dataset (supplied with this repository). To see all available networks run Part2.py --help. The RSLVQ implementation for it is found in rslvq2.py. Note that due to the lower number of features the complex networks may not be used with the MNIST dataset. To use the dogbreed dataset it needs to be correctly set up as described above.

common.py contains methods for loading the dataset used by both parts. convert.pl can be used to convert the output of Part1.py to CSV. Example: perl convert.pl < output.txt > output.csv. mnist.mat is the MNIST dataset used in part 2.